Legal claims defining the scope of protection. Each claim is shown in both the original legal language and a plain English translation.
1. An apparatus for computing, comprising: a computer processor comprising at least one central processing unit (“CPU”); a separability module to be operated by the computer processor to determine if a data class in a plurality of data classes is separable, wherein to determine if the data class is separable, the separability module is to determine an average intra-class similarity within each class in the plurality of data classes, an inter-class similarity across all data classes in the plurality of data classes, and is to determine separability of the data class based on the average intra-class similarity relative to the inter-class similarity; and an output and implementation module to be operated by the computer processor, to output a result of the separability of the data class to a data collector, wherein the data collector is to adapt data collection based at least in part on the result of the separability of the data class; wherein to determine the inter-class similarity across all data classes, the separability module is to determine, for a pair of data classes in the plurality of data classes, an average inter-class similarity, wherein the average inter-class similarity is determined either according to a similarity value for each signal in a first class in the pair of data classes relative to each signal in a second class in the pair of data classes and an average of such similarity values, or a similarity value for each signal in a first class in the pair of data classes relative to the average signal in a second class in the pair of data classes and an average of such similarity values, wherein the separability module is to fill a set of off-diagonal slots of a class separability matrix with the inter-class similarity for each pair of data classes in the plurality of classes, wherein the separability module is to, for each row in the class separability matrix, divide each off-diagonal slot in the row by a diagonal slot in the row and replace each off-diagonal slot with the result thereof; and wherein the diagonal slots of the class separability matrix are filled with the average intra-class similarity within each class.
This invention relates to a computing apparatus designed to evaluate the separability of data classes, addressing challenges in data classification and adaptive data collection. The apparatus includes a computer processor with at least one CPU and two key modules: a separability module and an output and implementation module. The separability module assesses whether a data class within a set of data classes is separable by calculating the average intra-class similarity for each class and the inter-class similarity across all classes. Separability is determined by comparing the intra-class similarity to the inter-class similarity. The inter-class similarity is computed for each pair of data classes by either averaging similarity values between individual signals in the classes or between signals in one class and the average signal of the other class. These values populate the off-diagonal slots of a class separability matrix, while the diagonal slots contain the average intra-class similarity. The matrix is normalized by dividing each off-diagonal slot by the corresponding diagonal slot. The output and implementation module then provides the separability results to a data collector, which adjusts data collection strategies based on these findings. This system enhances data classification accuracy and optimizes data collection processes by dynamically adapting to the separability of data classes.
2. The apparatus according to claim 1 , wherein to determine the average intra-class similarity within each class, the separability module is to determine, for each class, either an intra-class similarity value for all pairs of signals within a then-current class and an average of the intra-class similarity value for all pairs of signals within the then-current class, or an average intra-class value of all signals within a then-current class, a similarity of each signal in the then-current class relative to the average intra-class value of all signals within the then-current class, and an average of the similarity of each signal relative to the average intra-class value for each class, wherein the separability module is to further fill the set of diagonal slots of the class separability matrix with the average intra-class similarity within each class.
This invention relates to a system for evaluating signal separability in classification tasks, particularly in scenarios where signals are grouped into classes. The problem addressed is the need to assess how well signals within the same class are similar to each other, which is crucial for improving classification accuracy. The system includes a separability module that computes intra-class similarity metrics to quantify how distinct or overlapping signals are within each class. The separability module calculates an average intra-class similarity for each class using one of two methods. In the first method, it computes a similarity value for every pair of signals within a class and then averages these values. In the second method, it first determines an average signal representative of the class, then measures the similarity of each individual signal to this average, and finally averages these individual similarities. The results are used to populate a class separability matrix, where the diagonal entries represent the average intra-class similarity for each class. This matrix helps evaluate how well signals are grouped, aiding in tasks like feature selection, clustering, or classification refinement. The approach ensures that signals within the same class are consistently similar, improving the reliability of classification models.
3. The apparatus according to claim 2 , wherein the separability module is to determine a pair of data classes to be inseparable from one another when an off-diagonal slot at an intersection of the pair of data classes in the class separability matrix has a value greater than an inter-class threshold and is to either combine the pair of data classes into one class in the plurality of data classes for a machine learning problem or drop one of the pair of data classes for the machine learning problem.
This invention relates to machine learning systems that process data classes to improve classification performance. The problem addressed is the presence of inseparable or overlapping data classes, which can degrade model accuracy. The apparatus includes a separability module that evaluates class separability using a class separability matrix. This matrix contains off-diagonal slots representing the separability between pairs of data classes. When an off-diagonal slot exceeds an inter-class threshold, the module identifies the pair of classes as inseparable. The system then either merges the inseparable classes into a single class or removes one of the classes from the dataset. This adjustment ensures that the remaining classes are sufficiently distinct, enhancing the machine learning model's ability to classify data accurately. The separability matrix is derived from a feature space analysis, where class separability is quantified based on feature distributions or other statistical measures. The apparatus may also include a feature selection module that identifies relevant features for class separation, further refining the dataset. By dynamically adjusting the class structure, the system optimizes the input data for machine learning tasks, improving model performance and reliability.
4. The apparatus according to claim 2 , wherein the separability module is to determine the data class to be highly variable when a diagonal slot of the data class in the set of diagonal slots has a value less than an intra-class threshold and wherein the output and implementation module is to output the result that the data class is highly variable and is to remove the data class from the plurality of data classes for a machine learning problem.
This invention relates to data classification and machine learning, specifically addressing the challenge of identifying and handling highly variable data classes that may negatively impact model performance. The apparatus includes a separability module and an output and implementation module. The separability module evaluates the variability of data classes by analyzing diagonal slots in a set of diagonal slots associated with each class. If the value in a diagonal slot for a given data class falls below an intra-class threshold, the separability module classifies that data class as highly variable. The output and implementation module then generates a result indicating the high variability of the identified class and removes it from the set of data classes used in the machine learning problem. This process ensures that only stable, well-defined data classes are retained for training, improving the reliability and accuracy of the machine learning model. The apparatus may also include a data acquisition module to collect input data and a data processing module to preprocess the data before analysis. The invention aims to enhance machine learning performance by systematically filtering out problematic data classes that could introduce noise or bias into the model.
5. The apparatus according to claim 1 , wherein the computer processor further comprises a hardware accelerator encoded with a logic to perform a comparison; wherein the logic is used by the separability module to determine if the data class in a plurality of data classes is separable; wherein the logic to perform the comparison is executed at least in part by a set of artificial neurons of the hardware accelerator; and wherein pairs of signals are loaded in the artificial neurons at least in part to determine if the data class in the plurality of data classes is separable.
This invention relates to a hardware-accelerated system for determining the separability of data classes in machine learning or data processing applications. The problem addressed is the computational inefficiency of traditional software-based methods for assessing whether data classes can be distinguished, which is critical for tasks like classification, clustering, and feature selection. The solution involves a specialized hardware accelerator integrated with a computer processor, designed to accelerate separability analysis using artificial neurons. The hardware accelerator includes logic circuits configured to perform comparisons between data classes. These comparisons are executed at least partially by artificial neurons within the accelerator, which process pairs of signals representing data features or class attributes. The accelerator's logic determines whether a given data class in a set of classes is separable from others, meaning it can be distinctly identified based on the input data. This separability assessment is used by a separability module to optimize subsequent data processing tasks, such as training machine learning models or refining feature representations. By offloading separability computations to dedicated hardware, the system reduces processing time and energy consumption compared to software-based approaches. The use of artificial neurons in the accelerator enables parallel processing of signal pairs, further enhancing efficiency. This approach is particularly useful in high-performance computing environments where rapid data analysis is required.
6. A computer implemented method, comprising: determining if a data class in a plurality of data classes is separable, wherein determining if the data class is separable comprises determining an average intra-class similarity within each data class in the plurality of data classes, an inter-class similarity across all data classes of the plurality of data classes, and determining separability of the data class based on the average intra-class similarity relative to the inter-class similarity; and adapting a data collection based at least in part on a result of determining if the data class in the plurality of data classes is separable; wherein determining the inter-class similarity across all data classes comprises determining, for a pair of data classes in the plurality of data classes, an average inter-class similarity, wherein determining the average inter-class similarity comprises either determining a similarity value for each signal in a first class in the pair of data classes relative to each signal in a second class in the pair of data classes and an average of such similarity values, or determining a similarity value for each signal in a first class in the pair of data classes relative to the average signal in a second class in the pair of data classes and an average of such similarity values, further comprising filling a set of off-diagonal slots of a class separability matrix with the inter-class similarity for each pair of data classes in the plurality of classes, further comprising, for each row in the class separability matrix, dividing each off-diagonal slot in the row by a diagonal slot in the row and replacing each off-diagonal slot with the result thereof; wherein the diagonal slots of the class separability matrix are filled with the average intra-class similarity within each class.
This invention relates to a computer-implemented method for evaluating and adapting data collection based on the separability of data classes. The method addresses the problem of determining whether distinct data classes can be reliably distinguished from one another, which is critical for tasks such as classification, clustering, and data analysis. The approach involves assessing the separability of data classes by calculating an average intra-class similarity within each class and an inter-class similarity across all classes. The intra-class similarity measures how similar data points are within the same class, while the inter-class similarity measures how similar data points are between different classes. The separability is determined by comparing these two metrics. For inter-class similarity, the method computes pairwise similarities between signals in different classes, either by comparing each signal in one class to every signal in another class or by comparing each signal in one class to the average signal in another class. These values are then averaged and used to populate a class separability matrix. The diagonal slots of the matrix represent the average intra-class similarity for each class, while the off-diagonal slots represent the inter-class similarity. The off-diagonal values are normalized by dividing each by the corresponding diagonal value in the same row. The results of this analysis are used to adapt the data collection process, ensuring that the collected data is sufficiently distinct for accurate classification or analysis. This method improves data quality by dynamically adjusting collection parameters based on separability metrics.
7. The method according to claim 6 , wherein determining the average intra-class similarity within each data class comprises determining, for each data class, either an intra-class similarity value for all pairs of signals within a then-current class and an average of the intra-class similarity value for all pairs of signals within the then-current class, or an average intra-class value of all signals within a then-current class, a similarity of each signal in the then-current class relative to the average intra-class value of all signals within the then-current class, and an average of the similarity of each signal relative to the average intra-class value for each class, further comprising filling the set of diagonal slots of the class separability matrix with the average intra-class similarity within each data class.
This technical summary describes a method for evaluating data class separability in a classification system. The method addresses the challenge of accurately assessing how distinct different data classes are by measuring intra-class similarity, which helps improve classification performance. The method involves determining the average intra-class similarity for each data class. For a given class, this is done in one of two ways. First, it may calculate an intra-class similarity value for every pair of signals within the class and then compute the average of these values. Alternatively, it may calculate an average intra-class value for all signals in the class, determine the similarity of each signal relative to this average, and then average these individual similarities. The resulting average intra-class similarity is then used to fill the diagonal slots of a class separability matrix, which helps quantify how well the classes are separated. This approach ensures that the separability matrix accurately reflects the internal consistency of each class, aiding in better classification decisions.
8. The method according to claim 7 , further comprising determining the data class to be highly variable when a diagonal slot of the data class in the set of diagonal slots has a value less than an intra-class threshold and removing the data class from the plurality of data classes for a machine learning problem.
This invention relates to data processing for machine learning, specifically addressing the challenge of handling highly variable data classes that may degrade model performance. The method involves analyzing a set of diagonal slots, where each slot corresponds to a data class and contains a value representing variability within that class. A data class is identified as highly variable if its corresponding diagonal slot value falls below an intra-class threshold. Once identified, such highly variable data classes are removed from the dataset to improve the quality of the input for machine learning tasks. The method builds on a prior step of generating a diagonal matrix from a covariance matrix, where the diagonal slots represent intra-class variance. By filtering out these problematic classes, the approach aims to enhance model accuracy and stability. The technique is particularly useful in scenarios where certain data classes exhibit excessive variability, which could otherwise lead to overfitting or unreliable predictions. The removal process ensures that only more consistent and representative data classes are retained for training, thereby optimizing the learning process.
9. The method according to claim 7 , further comprising determining a pair of data classes to be inseparable from one another when the off-diagonal slot at an intersection of the pair of data classes in the class separability matrix has a value greater than an inter-class threshold and either combining the pair of data classes into one class in the plurality of data classes for a machine learning problem or dropping one of the pair of data classes for the machine learning problem.
This invention relates to machine learning systems that process data classes to improve classification performance. The problem addressed is the presence of inseparable or overlapping data classes, which can degrade model accuracy. The solution involves analyzing a class separability matrix to identify pairs of data classes that cannot be reliably distinguished from one another. The matrix contains off-diagonal values representing the separability between class pairs. When an off-diagonal value exceeds a predefined inter-class threshold, the system determines that the corresponding data classes are inseparable. In response, the system either merges the inseparable classes into a single class or removes one of the classes from the dataset. This adjustment ensures that the machine learning model is trained on a dataset with well-separated classes, enhancing classification accuracy. The method is particularly useful in scenarios where overlapping or ambiguous class boundaries exist, such as in image recognition, natural language processing, or other domains requiring precise categorization. By dynamically refining the class structure, the system optimizes the training process and improves model performance.
10. The method according to claim 6 , wherein determining is performed at least in part with a hardware accelerator encoded with a logic to perform a comparison, the hardware accelerator having a set of artificial neurons; and wherein determining comprises loading pairs of signals in the artificial neurons and using the artificial neurons at least in part to determine if the data class in the plurality of data classes is separable.
This invention relates to a method for determining whether data classes are separable using a hardware accelerator with artificial neurons. The method addresses the challenge of efficiently classifying data into distinct categories by leveraging specialized hardware to accelerate the separation analysis. The hardware accelerator is encoded with logic to perform comparisons and includes a set of artificial neurons. These neurons are used to process pairs of signals, enabling the system to assess whether the data classes can be distinguished from one another. The artificial neurons facilitate rapid and accurate determination of separability by evaluating the relationships between data points in different classes. This approach improves computational efficiency and accuracy compared to traditional software-based methods, particularly in applications requiring real-time or high-throughput data classification. The hardware accelerator's design allows for parallel processing of multiple data pairs, further enhancing performance. The method is applicable in fields such as machine learning, pattern recognition, and data analytics, where distinguishing between data classes is critical.
11. An apparatus for computing, comprising: means to determine if a data class in a plurality of data classes is separable, wherein means to determine if the data class is separable comprises means to determine an average intra-class similarity within each data class in the plurality of data classes, means to determine an inter-class similarity across all data classes of the plurality of data classes, and means to determine separability of the data class based on the average intra-class similarity relative to the inter-class similarity; means to adapt a data collection based at least in part on a result obtained from the means to determine if the data class in the plurality of data classes is separable; wherein means to determine the inter-class similarity across all data classes comprises means to determine, for a pair of data classes in the plurality of data classes, an average inter-class similarity, wherein means to determine the average inter-class similarity comprises either means to determine a similarity value for each signal in a first class in the pair of data classes relative to each signal in a second class in the pair of data classes and an average of such similarity values, or means to determine a similarity value for each signal in a first class in the pair of data classes relative to the average signal in a second class in the pair of data classes and an average of such similarity values, further comprising means to fill a set of off-diagonal slots of a class separability matrix with the inter-class similarity for each pair of data classes in the plurality of classes and, for each row in the class separability matrix, means to divide each off-diagonal slot in the row by a diagonal slot in the row and replace each off-diagonal slot with a result thereof; wherein the diagonal slots of the class separability matrix are filled with the average intra-class similarity within each class.
The apparatus is designed for computing separability of data classes in a dataset, addressing challenges in data classification where distinguishing between classes is difficult due to overlapping or ambiguous features. The system evaluates separability by calculating an average intra-class similarity for each data class, representing how similar data points are within the same class, and an inter-class similarity across all classes, representing how similar data points are between different classes. Separability is determined by comparing these values—higher intra-class similarity relative to inter-class similarity indicates better separability. The inter-class similarity is computed for each pair of data classes by either comparing each signal in one class to every signal in another class or comparing each signal in one class to the average signal of another class. The results are averaged to produce an inter-class similarity value. These values populate the off-diagonal slots of a class separability matrix, while the diagonal slots contain the average intra-class similarity for each class. The off-diagonal values are then normalized by dividing each by the corresponding diagonal value, refining the separability assessment. The apparatus adapts data collection based on the separability results, potentially improving classification accuracy by identifying and addressing poorly separable classes. This approach enhances data-driven decision-making in applications like machine learning, signal processing, and pattern recognition.
12. The apparatus according to claim 11 , wherein means to determine the average intra-class similarity within each data class comprises means to determine, for each data class, either an intra-class similarity value for all pairs of signals within a then-current class and an average of the intra-class similarity value for all pairs of signals within the then-current class, or an average intra-class value of all signals within a then-current class, a similarity of each signal in the then-current class relative to the average intra-class value of all signals within the then-current class, and an average of the similarity of each signal relative to the average intra-class value for each class, further comprising means to fill the set of diagonal slots of the class separability matrix with the average intra-class similarity within each data class.
This invention relates to a system for analyzing data classes by evaluating intra-class similarity and inter-class separability. The problem addressed is the need to accurately assess how distinct or similar data classes are within a dataset, particularly in applications like pattern recognition, clustering, or classification tasks. The apparatus includes a mechanism to compute the average intra-class similarity for each data class. For a given class, this involves either calculating a similarity value for every pair of signals within the class and averaging those values, or computing an average signal for the class, determining the similarity of each signal to this average, and then averaging those similarities. The resulting average intra-class similarity is then used to populate the diagonal elements of a class separability matrix, which helps quantify how well the classes are separated from one another. This approach ensures that the matrix accurately reflects both intra-class cohesion and inter-class distinctions, improving the reliability of subsequent analysis or decision-making processes. The system enhances the precision of data classification by providing a robust metric for evaluating class structure.
13. The apparatus according to claim 12 , further comprising means to determine the data class to be highly variable when a diagonal slot of the data class in the set of diagonal slots has a value less than an intra-class threshold and further comprising means to remove the data class from the plurality of data classes for a machine learning problem.
This invention relates to data processing for machine learning, specifically addressing the challenge of handling highly variable data classes that can degrade model performance. The apparatus includes a system that analyzes data classes to identify those with excessive variability, which can introduce noise or instability in machine learning tasks. The system evaluates each data class by examining its corresponding diagonal slot in a set of diagonal slots, where each slot represents a statistical measure of variability. If the value in the diagonal slot for a given data class falls below an intra-class threshold, the system classifies that data class as highly variable. Once identified, the apparatus removes the highly variable data class from the set of data classes used in the machine learning problem, thereby improving model robustness and accuracy. The apparatus may also include means to compute or update the intra-class threshold dynamically based on the distribution of diagonal slot values across all data classes. This ensures adaptive filtering of problematic data classes as the dataset evolves. The invention is particularly useful in scenarios where data quality is inconsistent, such as in real-world datasets with outliers or imbalanced classes.
14. The apparatus according to claim 12 , further comprising means to determine a pair of data classes to be inseparable from one another when an off-diagonal slot at an intersection of the pair of data classes in the class separability matrix has a value greater than an inter-class threshold and further comprising means to either combine the pair of data classes into one class in the plurality of data classes for a machine learning problem or drop one of the pair of data classes for the machine learning problem.
This invention relates to machine learning systems that process data classes to improve classification accuracy. The problem addressed is the presence of inseparable data classes, which can degrade model performance by introducing noise or redundancy. The apparatus includes a class separability matrix that quantifies the separability between pairs of data classes. When an off-diagonal slot in this matrix exceeds an inter-class threshold, the system identifies the corresponding pair of data classes as inseparable. To resolve this, the apparatus either merges the inseparable classes into a single class or removes one of the classes from the dataset. This ensures that the machine learning model operates on a more distinct set of classes, enhancing classification accuracy. The inter-class threshold is a configurable parameter that determines the strictness of the separability criterion. The apparatus may also include means to compute the class separability matrix, which involves evaluating statistical or distance-based metrics between class distributions. The solution is particularly useful in scenarios where overlapping or ambiguous class boundaries exist, such as in image recognition, natural language processing, or other multi-class classification tasks. By dynamically adjusting the class structure, the system optimizes the input data for machine learning algorithms, leading to improved generalization and reduced overfitting.
15. The apparatus according to claim 11 , wherein the means to determine comprises a hardware accelerator encoded with a logic to perform a comparison; and wherein the apparatus further comprises an internal sensor subsystem controlled at least in part by the hardware accelerator and further comprising means to receive an identification of the plurality of data classes and means to collect with the internal sensor subsystem a set of signal data comprising the plurality of data classes, wherein the set of signal data have a common set of units.
This invention relates to an apparatus for processing and classifying sensor data using a hardware accelerator. The apparatus addresses the challenge of efficiently categorizing diverse sensor data into predefined classes while ensuring consistent measurement units. The hardware accelerator is encoded with logic to perform comparisons between sensor data and predefined data classes. The apparatus includes an internal sensor subsystem, which is at least partially controlled by the hardware accelerator. The system is capable of receiving an identification of multiple data classes and collecting a set of signal data from the internal sensor subsystem, where all collected data shares a common set of units. This ensures uniformity in data processing and classification. The hardware accelerator's logic enables rapid and accurate classification of sensor data, improving efficiency in applications requiring real-time or high-throughput data analysis. The internal sensor subsystem may include various sensors, and the collected data is processed to ensure it conforms to the specified units, facilitating seamless integration with the classification logic. This approach enhances the reliability and speed of data classification in sensor-based systems.
16. One or more non-transitory computer-readable media comprising instructions that cause a computer device, in response to execution of the instructions by a processor of the computer device, to: determine if a data class in a plurality of data classes is separable, wherein determine if the data class is separable comprises determine an average intra-class similarity within each data class in the plurality of data classes, determine an inter-class similarity across all data classes in the plurality of data classes, and determine separability of the data class based on the average intra-class similarity relative to the inter-class similarity; and adapt a data collection based at least in part on a result obtained from the determination if the data class in the plurality of data classes is separable; wherein determine the inter-class similarity across all data classes comprises determine, for a pair of data classes in the plurality of data classes, an average inter-class similarity, wherein determine the average inter-class similarity comprises either determine a similarity value for each signal in a first class in the pair of data classes relative to each signal in a second class in the pair of data classes and an average of such similarity values, or determine a similarity value for each signal in a first class in the pair of data classes relative to the average signal in a second class in the pair of data classes and an average of such similarity values, and where the instructions are further to cause the processor to fill a set of off-diagonal slots of a class separability matrix with the inter-class similarity for each pair of data classes in the plurality of classes and, for each row in the class separability matrix, divide each off-diagonal slot in the row by the diagonal slot in the row and replace each off-diagonal slot with the result thereof wherein the diagonal slots of the class separability matrix are filled with the average intra-class similarity within each class.
The invention relates to a system for analyzing and adapting data collection based on the separability of data classes. The problem addressed is the need to assess whether distinct data classes can be reliably distinguished from one another, which is critical for tasks such as classification, clustering, and data-driven decision-making. The system evaluates separability by computing an average intra-class similarity for each data class, representing how similar data points within the same class are to each other. It also calculates an inter-class similarity, which measures how similar data points from different classes are. The inter-class similarity is determined either by comparing each signal in one class to every signal in another class or by comparing each signal in one class to the average signal of another class. These values are used to populate a class separability matrix, where diagonal slots contain the intra-class similarity for each class, and off-diagonal slots contain the inter-class similarity between pairs of classes. The matrix is normalized by dividing each off-diagonal slot by the corresponding diagonal slot, providing a relative measure of separability. Based on this analysis, the system adapts data collection strategies to improve class separability, ensuring more distinct and reliable data classes for subsequent processing.
17. The computer-readable media according to claim 16 , wherein determine the average intra-class similarity within each data class comprises determine, for each data class, either i) an intra-class similarity value for all pairs of signals within a then-current class and an average of the intra-class similarity value for all pairs of signals within the then-current class, or ii) an average intra-class value of all signals within a then-current class, a similarity of each signal in the then-current class relative to the average intra-class value of all signals within the then-current class, and an average of the similarity of each signal relative to the average intra-class value for each class, wherein the instructions are further to cause to the processor to fill the set of diagonal slots of the class separability matrix with the average intra-class similarity within each data class.
This invention relates to a method for analyzing signal data to improve classification accuracy by evaluating intra-class similarity within data classes. The problem addressed is the need to assess how similar signals within the same class are to each other, which is crucial for determining class separability and improving classification performance. The invention involves computing intra-class similarity metrics for each data class, which can be done in two ways. First, it may calculate a similarity value for every pair of signals within a class and then average those values. Alternatively, it may compute an average signal for the class, measure the similarity of each signal to this average, and then average those individual similarities. These computed intra-class similarity values are then used to fill the diagonal slots of a class separability matrix, which helps in evaluating how distinct or overlapping the classes are. This approach enhances the ability to distinguish between different classes by quantifying internal consistency within each class, thereby improving classification algorithms. The method is particularly useful in applications where signal data must be accurately categorized, such as in pattern recognition, machine learning, and data clustering tasks.
18. The computer-readable media according to claim 17 , further comprising determine the data class to be highly variable when a diagonal slot of the data class in the set of diagonal classes has a value less than an intra-class threshold and remove the data class from the plurality of data classes for a machine learning problem.
The invention relates to data classification and preprocessing for machine learning, specifically addressing the challenge of handling highly variable data classes that may negatively impact model performance. The system identifies and removes data classes that exhibit excessive variability, which can lead to noise or instability in machine learning models. The process involves analyzing a set of diagonal classes, where each class is represented by a diagonal slot containing a variability metric. If the value in this slot falls below an intra-class threshold, the data class is classified as highly variable and excluded from the dataset used for training or inference. This ensures that only stable, representative data classes are retained, improving model accuracy and robustness. The method is particularly useful in scenarios where datasets contain outliers or classes with inconsistent patterns, such as in image recognition, natural language processing, or financial forecasting. By dynamically filtering out problematic classes, the system enhances the reliability of machine learning workflows.
19. The computer-readable media according to claim 17 , wherein the instructions further cause the processor to determine a pair of data classes to be inseparable from one another when an off-diagonal slot at the intersection of the pair of data classes in the class separability matrix has a value greater than an inter-class threshold.
The invention relates to data classification systems that analyze separability between data classes using a class separability matrix. The problem addressed is determining whether pairs of data classes are inseparable, which is critical for improving classification accuracy and reducing misclassification errors. The system generates a class separability matrix where each off-diagonal slot represents the separability between two distinct data classes. When the value in an off-diagonal slot exceeds a predefined inter-class threshold, the system identifies the corresponding pair of data classes as inseparable. This determination helps in refining classification models by highlighting data classes that cannot be reliably distinguished, allowing for adjustments such as feature engineering, data augmentation, or model retraining. The inter-class threshold is a configurable parameter that defines the minimum separability required to consider two classes distinguishable. The system processes the class separability matrix to flag inseparable class pairs, enabling downstream improvements in classification performance. This approach is particularly useful in machine learning applications where accurate class distinction is essential, such as medical diagnosis, fraud detection, and quality control.
20. The computer-readable media according to claim 16 , wherein to determine separability of the data class based on the average intra-class similarity relative to the inter-class similarity determines that the pair of data classes are inseparable and further comprising either combine the pair of data classes into one class in the plurality of data classes for a machine learning problem or drop one of the pair of data classes for the machine learning problem.
This invention relates to machine learning systems that optimize data class separation for improved model performance. The problem addressed is the presence of data classes that are too similar to be meaningfully distinguished, which can degrade model accuracy and efficiency. The solution involves analyzing intra-class (within-class) and inter-class (between-class) similarity metrics to assess class separability. The system evaluates whether a pair of data classes can be reliably distinguished based on their average intra-class similarity relative to inter-class similarity. If the classes are determined to be inseparable, the system takes corrective action. One approach is to merge the inseparable classes into a single, more distinct class, reducing redundancy and improving model training. Alternatively, the system may drop one of the inseparable classes entirely, simplifying the dataset and focusing the model on more distinguishable categories. This method ensures that the machine learning model operates on a dataset where classes are sufficiently distinct, enhancing classification accuracy and computational efficiency. The approach is particularly useful in scenarios where overlapping or ambiguous class boundaries exist, such as in image recognition, natural language processing, or other domains with complex feature spaces. By dynamically adjusting the class structure, the system adapts to the inherent separability of the data, leading to more robust and interpretable machine learning outcomes.
21. The computer-readable media according to claim 16 , wherein to determine comprises to use a hardware accelerator encoded with a logic to perform a comparison; and wherein the hardware accelerator comprises a set of artificial neurons; and wherein to determine comprises to load pairs of signals in the artificial neurons to determine if the data class in the plurality of data classes is separable.
This invention relates to a hardware-accelerated system for classifying data using artificial neurons. The problem addressed is the computational inefficiency of traditional software-based classification methods, particularly when processing large datasets with multiple classes. The solution involves a hardware accelerator specifically designed to perform rapid comparisons between data signals using artificial neurons. The hardware accelerator is encoded with logic to evaluate whether data classes are separable by loading pairs of signals into the artificial neurons. The artificial neurons function as processing units that compare input signals to determine class separability, enabling faster and more efficient classification compared to conventional software implementations. The system is particularly useful in applications requiring real-time or high-throughput data classification, such as machine learning, pattern recognition, and automated decision-making systems. The hardware accelerator's specialized design reduces latency and improves energy efficiency by offloading classification tasks from general-purpose processors. The invention leverages the parallel processing capabilities of artificial neurons to enhance performance, making it suitable for deployment in edge devices, data centers, and other computing environments where speed and efficiency are critical.
Unknown
August 25, 2020
Browse 5M+ US patents with plain-English claim translations and AI-generated analysis.